The Estimate of Land Surface Temperature Components for Soil and Vegetation Using the MODIS Dataset and an Ensemble Kalman Filter – Based Assimilation Approach in a Heterogeneous Mediterranean Ecosystem

Heterogeneous Mediterranean ecosystems are landscapes with scattered trees dispersed throughout a grassy matrix, and long-time series of land surface temperature (LST) images with high spatial and high temporal resolutions are needed to model their dynamics of the vegetation. Data assimilation systems have been developed for guiding the model with observations towards optimal solutions, and can be useful in the case of operational prediction approaches for reducing the uncertain model parameterization. For operational data assimilation approaches in heterogenous ecosystems there is the need of long-time series of LST images with high spatial and high temporal resolutions.  This is currently difficult because of the spatial-temporal trade-off associated with satellite-based observations of LST, and there is a need of obtaining long time series of high-spatial and high-temporal resolution thermal data. This prompted us to propose a novel downscaling procedure that used MOD11A1 and MYD11A1 (~1000 m spatial resolution) as source data, and a coarse (~ 1000 m) and a fine (~30 m) annual estimation of the NDVI as ancillary. The approach, tested in a ecosystem in Sardinia (Italy), supplied by an eddy-covariance station, led to the creation of ~7700 maps of LST (30 m) covering the years 2000-2022. A first validation was done by comparing 19 years of ground-based data with the LST estimates from satellites, while the second validation was performed spatially by comparing MODIS downscaled maps with ASTER (90 m) and LANDSAT (100 m) scenes. The approach was able to reduce the spatial scale of MODIS LST observations by maintaining their original time frequency. The use of the LST observations from MODIS using the downscaling approach allowed merging the LST data from remote sensors and the LSM optimally for predicting accurately grass and tree LST in the assimilation approach. A sensitivity analysis of the data assimilation approach to assimilation time interval demonstrated that the use of the MODIS time interval of acquisition (i.e., ~12 hours) allowed to obtain accurate results.